Evaluating probabilistic deep learning methods for uncertainty quantification of temperature downscaling

Abstract Deep learning (DL) has emerged as a promising tool for downscaling coarse-resolution climate data to high-resolution outputs, enabling improved regional climate predictions. A critical aspect of DL-based downscaling is the incorporation of uncertainty quantification (UQ), which enhances the interpretability and reliability of predictions—key factors for climate risk assessment and decision-making. This study develops a DL model to downscale 2 m temperature across the contiguous United States using reanalysis datasets. We systematically evaluate three epistemic UQ methods—deep ensembles (DEns), Monte Carlo dropout (MCD), and Flipout—based on their probabilistic accuracy, downscaling performance, sensitivity to geographical features, and computational efficiency. Results indicate that MCD generally outperforms Flipout and DEns in terms of calibration and downscaling accuracy. However, DEns demonstrate lower calibration errors in coastal regions, indicating its higher confidence within these areas. Flipout, in contrast, is more sensitive to elevation gradients and exhibits higher calibration errors in mountainous regions. Hence, the choice of UQ method for this task depends on the specific requirements of the application. For applications that prioritize overall calibration, downscaling accuracy, and computational efficiency, MCD is a strong candidate. These findings highlight the importance of selecting UQ methods based on application-specific requirements, such as geographical context and computational constraints. By addressing the trade-offs between UQ methods, this study provides actionable insights for improving the reliability, scalability, and utility of DL-based downscaling in climate science.